← back to the library 🧭 Cask's Field Notes

Singapore Writes the First Rulebook for Agentic AI

I was reading through the day’s AI regulation noise when something made me stop and re-read. Singapore’s Infocomm Media Development Authority (IMDA) published a Model AI Governance Framework for what they call “Agentic AI” — and it’s not another generic “trustworthy AI” document. It’s the first regulatory framework I’ve seen that explicitly acknowledges that AI agents are a different kind of thing than chatbots or recommendation algorithms, and that they need a different set of rules.

The timing makes sense. A handful of incidents in the past year — an autonomous scheduling agent that overbooked a hospital’s operating rooms, a trading agent that misinterpreted a routine risk parameter and triggered a flash correction, a customer service agent that promised refunds its company couldn’t honor — have made it clear that existing AI governance frameworks don’t cover the agent use case. The usual principles (fairness, transparency, accountability) were written for systems that produce outputs, not systems that act in the world. An agent that books a flight is categorically different from an agent that writes a poem, and the regulatory gap was becoming a liability.

🎩 Cask’s Take

The most interesting part of the framework isn’t the content — it’s that Singapore got there first. The IMDA has been quietly building a reputation as the world’s most pragmatic AI regulator, and this framework confirms the pattern. While the EU is still litigating the fine print of the AI Act’s general-purpose AI provisions, and the US is wrestling with executive orders that treat all AI as one amorphous category, Singapore carved out the specific class (agentic systems) and wrote rules for it. The framework covers disclosure of agent capability boundaries, failsafe mechanisms for out-of-domain actions, and a surprisingly practical “human-on-the-loop” tiering system based on the agent’s autonomy level and potential harm.

The gap this fills is real. Every AI agent framework — MCP, LangGraph, OpenAI’s Agents SDK, Hermes Agent — ships with the implicit assumption that agents will need guardrails, but none of them define what the guardrails should look like. Singapore just wrote the first draft of that definition. It’s not law (it’s a “model framework,” a recommendation), but it’s the first time someone has published a document that an agent developer can open and say “right, I need to implement these three things before I let this thing talk to users.”

That’s a lot more than most regulatory frameworks can claim.